Introduction

Future changes of tropical cyclone (TC) activities in changing climate have always been a topic of great research interest. Landfalling TCs can bring fierce winds and heavy rainfall to coastal regions1, sometimes accompanied by abnormally high temperatures2, which cause severe economic losses and geological disasters. The western North Pacific (WNP), which is the most active basin for TCs, accounts for about 33% of global TC frequency. WNP TCs predominantly occur in the monsoon trough and are regulated by the large-scale environment and sea surface temperature (SST) forcings, including the tropical easterly jet3, El Niño/Southern Oscillation (ENSO)4,5, Pacific Decadal Oscillation (PDO)6, and so on7,8. Furthermore, TC frequency, track, and intensity have a key effect on landfall number, landfall intensity, and rainfall rate9,10.

As global warming continues, TC activity changes accordingly11,12,13,14, posing an increasing threat to coastal areas15,16. Recent studies have indicated that global TC activity, including the position where TCs occur and reach their lifetime maximum intensity (LMI), presents a poleward migration, especially over the WNP15,17,18,19,20. Meanwhile, some argue that this poleward shift has reversed since around 199921. The inconsistency among these results may be primarily attributed to the different datasets and investigation periods. Sharmila and Walsh17 linked the poleward migration to tropical expansion and the shift in the regional Hadley circulation under global warming. The La Niña-like cooling pattern of tropical SST has also contributed to the seasonal change in TC frequency and thus genesis location22. The phase change of PDO is also thought to play a dominant role23. Moreover, TC genesis and LMI locations have shifted westward toward the coast24,25, and the intensity of landfalling TCs has increased moderately16.

In recent years, numerous studies have endeavored to reproduce and project TC activity using meteorological models26,27. Several model projections suggest that the poleward migration observed over the WNP would continue in the future28,29,30,31. However, current climate models often underestimate TC occurrence compared to observations and tend to misidentify extratropical storms as tropical ones32. High-resolution climate models typically perform better than their lower-resolution counterparts33,34,35. As a result, the CMIP6 High-Resolution Model Intercomparison Project (HighResMIP) has aroused significant interest36,37, especially when compared to previous models with coarser horizontal resolutions. Several studies have evaluated the simulation of TCs by CMIP6-HighResMIP models38,39,40. However, they have primarily focused on TC frequency and track density, with less emphasis on TC genesis and LMI location, which are closely related to the poleward migration of TCs.

In this study, we employ the high-resolution climate models from CMIP6-HighResMIP to investigate the future projected poleward migration of TCs over the WNP, and to identify the possible driving factors. We assess the model performance based on the observations and introduce a revised constrained detection method. This approach markedly reduces the model errors a lot and substantially boosts our confidence in model projections.

Results

A revised constrained detection method

Figure 1a, b illustrates the annual cycles of TC genesis frequency and latitude, comparing observations with simulations from the HighResMIP models (see the “Methods” section). There are significant differences in the performance of the two tracking algorithms in terms of both genesis frequency and latitude, with TempestExtremes41 (referred to as TempExt) showing closer alignment to the observations. The reconstructed frequency peaks in the summer and autumn months, reflecting the single-peak characteristic observed across the twelve months. Variability is noted among the four models assessed. EC-Earth3P-HR exhibits the lowest frequency, which is much lower than observations. In contrast, HadGEM3-GC31-HM demonstrates the highest frequency, which surpasses the observed counts by 1–2 per month throughout the year. In general, the monthly frequency in winter and spring is unusually high, particularly with the TRACK algorithm42, reaching two to six times the frequency observed. This finding is consistent with previous studies38. Compared with the frequency, the genesis latitude simulated by the models is less satisfactory. In the TRACK algorithm, the result fails to capture the annual cycle. Notably, the latitude in winter and spring is abnormally high, surpassing even in summer. CNRM_TempExt and HadGEM3_TempExt demonstrate the best performance.

Fig. 1: Comparison of TC genesis frequency and latitude between observation and HighResMIP models.
figure 1

The climatological mean (1980–2010) annual cycle of a genesis frequency and b latitude over the WNP, from IBTrACS (black bars), the models using the TRACK tracker (solid lines) and the TempestExtremes tracker (dashed lines), as well as their MMEs (gray bars). The rightmost column represents the annual mean genesis frequency and latitude using the TRACK tracker (colored dots) and the TempestExtremes tracker (colored circles). c, d Same as in (a, b), but under the revised constrained detection method.

Full size image

Why does the TRACK algorithm show great inconsistency compared to the TempExt and observation results, especially in winter and spring? Considering that general circulation models (GCMs) may misidentify an extratropical storm as a tropical storm32, an analysis of the monthly genesis distribution for both trackers and each model during 1980–2010 was conducted. Supplementary Fig. 1a depicts the genesis distribution of CMCC using the TRACK tracker. In January, a number of storms occur between 20°N and 35°N, well above the 26 °C isotherm (black line), by ~10°N. This pattern is also observed in other months, including February, March, and December. Corroborated with the 200-hPa zonal wind in Supplementary Fig. 3, these storms, in fact, lie in the westerlies. Similar results occur among other models when tracked as the TRACK algorithm, but this is not the case with TempExt (Supplementary Figs. 1, 2). Therefore, the TRACK algorithm overestimated the frequency of TCs, especially in winter and spring, leading to abnormally high genesis latitude. Furthermore, the algorithm occasionally identifies storms over land, which is clearly unreasonable and needs elimination. Previous studies also made constraints using a uniform latitude threshold or duration limitation39,40. However, these methods have been ineffective in reducing the error caused by misidentifying extratropical storms as tropical storms. To address this issue, we put forward a revised constrained detection method based on the monthly location of the 26 °C SST isotherm and the westerlies. The genesis latitude threshold, which lies south of the westerly belt and north of the 26 °C isotherm, varies by season as follows: it is set at 20°N for December and January–March; 25°N for November, April, and May; 30°N for June and October; and 35°N for July–September. Meanwhile, the samples that appear over land are also removed. Figure 1c, d shows the climatological monthly mean formation frequency and latitude under the revised method. The frequency in winter and spring apparently approaches the observations more closely, and the multimodel ensemble means (MMEs) are largely consistent with the observations. EC-Earth3P-HR and HadGEM3-GC31-HM still exhibit the lowest and highest frequencies, respectively. The revised method led to a greater consistency among the models, with 1–3°overestimation compared with observed latitude. In general, both the TRACK and TempExt algorithms basically simulate the annual cycle of TCs fairly well over the WNP, and the errors are reduced considerably (also see in Supplementary Table 2). The reliability of the proposed constrained detection method is also evidenced by the apparent reduction in discrepancies regarding annual TC genesis latitude and frequency compared with observations, particularly noticeable in the TRACK algorithm (Supplementary Fig. 4).

Poleward migration of TCs

We have applied the revised constrained method to extract projected future TCs and compare them with the past (see the “Methods” section). Figure 2a presents the future changes in monthly and annual genesis latitude over the WNP. There is considerable variability among the models from January to May, with some demonstrating a latitude increase while others a decrease. From June to December, the results are more concentrated and appear mostly to increase, especially in August and September. The change in annual latitude has the highest consistency among models, showing an increase of about 1°. Each model and its MME present significant poleward migration during the TC peak season, and the results for MME_TRACK and MME_TempExt are 1.13° and 1.16°, respectively (Supplementary Fig. 5). Furthermore, according to TC genesis density, regions south of 15° present a significant decrease while those north of 15° show a less significant increase (Fig. 2). It is noted that a small region east of 120°E and south of 15°N appears a significant increase throughout the year in Fig. 2d, while this region does not appear in the TC peak season (Fig. 2c). This is probably due to the slight increase in frequency in this region during winter and spring (Fig. 2b). Additionally, frequency changes in all four models and their MMEs during the TC peak season indicate an insignificant reduction (Supplementary Fig. 5c). This implies that the reduction of TCs south of 15°N during the TC peak season is the main factor responsible for the increase in genesis latitude during the TC peak season and throughout the year20.

Fig. 2: Future projections of TC genesis latitude and frequency for each model and its MME.
figure 2

Future projections of monthly and annual a TC genesis latitude b genesis frequency. The black dots and gray dots represent the MME_TRACK and MME_TempExt values, respectively. c Future projections of JJASO TC genesis density. d Same as in (c) but for annual changes. Black dots in c and d denote values statistically significant at the 95% confidence level.

Full size image

Besides TC genesis position, future changes in LMI are also noteworthy. The simulated LMI latitude is a little higher than the observed, especially during the TC peak season, but captures the annual cycle fairly well (Supplementary Fig. 6). During the TC peak season, the LMI latitude of each model and its MME all indicate a poleward migration, while the results from other months show poor consistency. (Supplementary Figs. 5b and 7).

To sum up, both Genesis latitude and LMI latitude show poleward migration in the near future, with the latter less significant, similar to previous results20,43. The phenomenon of poleward migration is associated with changes in the TC intensity. In this study, for the high-resolution models used, classifying the intensity of storms by surface air pressure proves more accurate than by maximum wind speed. This is because these models do not effectively simulate strong wind speeds, leading to an underestimation of the frequency of intense storms44,45. Following the storm intensity categories used in Roberts et al.37, the percentage changes of TCs over the WNP are analyzed, as measured by both maximum wind speed and minimum mean sea level pressure (MSLP) (Supplementary Table 3; Supplementary Figs. 8 and 9). Some models indeed fail to simulate very intense TCs measured by both maximum wind speeds (Supplementary Fig. 8). Overall, the projected proportion of category 5 is likely to increase. Additionally, Lin et al. analyzed the potential intensity and found that poleward migration can partially restrain the intensification of future TCs due to global warming30. However, it is worth noting that the simulated TCs may be weaker than the observed because of the lower resolution26,46.

DGPI and large-scale environmental change

The dynamic genesis potential index (DGPI) presents the impacts of environmental changes on TC formation and shows an increasing (decreasing) trend after 1998 at high (low) latitudes, according to Daloz and Camargo47. To investigate the environmental factors affecting the poleward migration of TCs over the WNP, the future changes in DGPI have been calculated as depicted in Fig. 4a (see the “Methods” section). In the primary region of the WNP, there is a significant increase (decrease) north (south) of approximately 20°N, corresponding to the poleward shift in TC genesis density in Fig. 2. Additionally, the nearshore area of eastern China also shows an apparent increase, which may be related to the uneven warming of SST or the nearshore errors of some environmental data in the models. Two regions, designed as A (10°N–16°N, 138°E–159°E) and B (21°N–27°N, 142°E–165°E), with similar longitude and distinct latitude, are selected to analyze the relative contribution of the DGPI (Fig. 3b–d) due to their significant variation. The 500 hPa vertical velocity makes the largest contribution in both regions, followed by the 850 hPa absolute vorticity and the meridional gradient of the 500 hPa zonal wind. The vertical wind shear term also makes negative contributions to area B, which means that the slight increase in vertical wind shear is unfavorable for TCs.

Fig. 3: Future projections of JJASO dynamic genesis potential index (DGPI) and relative contributions of its four terms.
figure 3

a Future projections of DGPI over the WNP and relative contributions of its four terms over b area A (10°N–16°N,138°E–159°E), c area B (21°N–27°N,142°E–165°E), and d the difference between A and B. Black dots and * denote values statistically significant at the 95% confidence level.

Full size image

The spatial distributions of large-scale environmental future changes are illustrated in Fig. 4. The 850 hPa absolute vorticity and 500 hPa vertical velocity changes over the WNP are similar to the DGPI changes, especially in areas A and B (Fig. 4a, b). The vorticity increases significantly north of 20°N, with a cyclonic circulation in the lower level, which favors TC genesis. The easterly wind intensifies in the western equatorial Pacific, and the Walker circulation may strengthen accordingly, which is not conducive to the formation and development of TCs22. The main formation region of TCs presents a decrease in vertical velocity and relative humidity and an increase in outgoing longwave radiation (OLR) at lower latitudes, such as in area A, while it is opposite at higher latitudes, for instance, in area B (Fig. 4b, d, e). Meanwhile, vertical wind shear (VWS) increases significantly around 5°N–10°N and 160°W-180° and extends northwestward. Areas A and B are both located in the area with a slight VWS increase, indicating that it has little effect on the poleward shift of WNP TCs, which supports the DGPI analysis. Overall, the TC-favorable conditions over the WNP transfer from south to north indicate a poleward migration as well.

Fig. 4: Composite differences of JJASO large-scale circulations.
figure 4

Composite differences of a 850 hPa relative vorticity (contours, 10−5 s−1) and winds (vectors, m s−1), b 500 hPa vertical velocity (contours, Pa s−1) and winds (vectors, m s−1), c vertical wind shear, d 600-hPa relative humidity (contours, %), e outgoing long-wave radiation (contours, W m−2). Black dots denote values statistically significant at the 95% confidence level.

Full size image

Discussion

The poleward migration of TC-favorable conditions has been conducive to the recent phase transition of the Pacific Decadal Oscillation (PDO)23, the intensification of the Walker circulation22 and tropical expansion1. The future changes in the vertical velocity and the meridional components of divergent wind show an anomalous subsidence occurred around 5–15°N and a relatively weak uplift around 15–25°N (Fig. 5), which corresponds with the changes in the DGPI and the environmental fields shown in Figs. 3 and 4. This manifests a close link between the poleward migration and the ascending branch of the regional Hadley circulation17,43. The weakening and poleward shift of the ascending branch of the Hadley circulation may be associated with the relatively increased vertical wind shear and convective stability in the deep tropics17. Anjana and Kumar43 noted that TCs exhibit significantly higher co-variability with the boundaries of ascending regions than descending regions, and the gradient in SST is a key factor for TC genesis and the dynamics of the Hadley Circulation.

Fig. 5: Meridional-vertical structure of the regional Hadley circulation during JJASO.
figure 5

a Meridional-vertical distribution of the longitudinally averaged divergent meridional wind (unit: m s−1) and vertical pressure velocity (unit: 10−2 Pa s−1) (shown in vector) over the WNP (averaged between 105°E and 180°) during P1 (1980–2010). b Same as in (a), but for P2 (2020–2050). c Composite differences between P2 and P1. The shading over the same region shows the vertical pressure velocity. Negative (positive) values represent upward (downward) motion. Black dots denote values statistically significant at the 95% confidence level.

Full size image

Supplementary Fig. 10 shows a correlation map between observed TC genesis latitude and detrended SST from 1980 to 2010, which presents a La-Niña-like pattern and is similar to the interdecadal Pacific oscillation (IPO) negative phase, indicating that internal variability plays a crucial role23,44. But constant and stronger global warming may also intensify TC poleward migration19. Under the SSP5–8.5 scenario, future radiative forcing will reach 8.5 W m−2 in 2100 with high emissions48. Compared with the past SST trend, global warming is pronounced in the future, especially in the middle and high latitudes (Supplementary Fig. 11), and TC frequency also shows a decreasing trend11,49, as indicated in Supplementary Fig. 12. Thus, the future uneven warming of sea surface temperature under the SSP5–8.5 scenario is likely to exert a combined effect with internal variability on the poleward migration of TCs, thereby potentially weakening the effects of internal variability.

Besides the Pacific, there is also a significant positive correlation over the North Atlantic and East Indian Ocean50,51,52. Liu et al. found that the cold central Pacific Ocean, the warm tropical North Indian Ocean, and the relatively warm tropical North Atlantic Ocean might be related to the absence of TCs in July 2020 over the WNP53. The abnormal warming of the east Indian Ocean SST can stimulate an eastward Kelvin wave, regulating the circulation over the main genesis region of WNP TCs and weakening the East Asian summer monsoon, and thus modulating TC location and frequency50,54,55. However, the influence of interactions between three oceans on the poleward migration of TCs over the WNP warrants further study.

This study proposes a revised constrained detection method for CMIP6-HighResMIP models based on monthly SST and westerlies, which effectively reduces the error caused by the misidentification of extratropical cyclones. The model results under the revised method capture the annual cycle of the frequency and latitude of TCs over the WNP fairly well, and the performance in reproducing interannual TC genesis latitude and frequency are also improved, which greatly enhances our confidence in the model construction and utilization and provides a reference for future improvements in TC simulation and forecasting capabilities. The monthly TC genesis distribution of the models at the standard resolution (e.g. CNRM-CM6-1) is similar to at the high resolution, indicating that our constrained detection method is applicable to other models in the CMIP6-HighResMIP (Supplementary Fig. 13). Additionally, the genesis wind speeds of some storms are unusually higher than 22 m s−1 and even each 30 m s−1, which is unreasonable and needs further investigation and model improvement (Supplementary Fig. 14).

Under the revised method, we analyze future changes in TC activity using the atmosphere-only (uncoupled) models of CMIP6-HighResMIP between the present period (1980–2010) and future period (2020–2050 under scenario SSP5–8.5) and find that both the genesis and LMI positions show a poleward migration. The latter is less significant and is caused mainly by the reduction of TCs in lower latitudes. Correspondingly, the DGPI increases in higher latitudes and decreases in lower latitudes, to which vertical velocity makes the largest contribution and vertical wind shear the least. The weakening and poleward shift of the regional Hadley circulation over the WNP is consistent with the northward movement of the TC-favorable environment fields. Under constant uneven warming, internal variability and global warming probably have joint effects on TC poleward migration.

Methods

Observation data

We use observational TC data from the International Best Track Archive for Climate Stewardship version 4 (IBTrACSv4) at 6-h intervals56. Cyclogenesis is defined as TC surface maximum wind speed reaching 34 knots for the first time with a latitude <35°N. Monthly mean SST data are obtained from the United States National Oceanic and Atmospheric Administration (NOAA) Extended Reconstruction SST version 5 (ERSSTv5)57, with a spatial resolution of 2° × 2°.

Models

The European Union Horizon 2020 project PRIMAVERA (Process-based climate simulation: Advances in high-resolution modeling and European climate risk assessments; https://www.climateurope.eu/primavera/) has six different contributing global atmospheric models, each run comprises model resolutions at both a standard (typically 100 km) and a high resolution (toward 25 km). These models are HighResMIP members, and their runs follow a common protocol proposed by HighResMIP. The high resolution ones are selected for their better performance in our study. The main experiments of these HighResMIP models can be divided into three tiers. Tier 1 (that is highresSST-present”) and Tier 3 (that is highresSST-future) are uncoupled (atmosphere-only) simulations. Tier 2 is coupled ocean-atmosphere simulations, including two parts: hist-1950 and highres-future. We utilize highresSST-present and highresSST-future to evaluate model performance and analyze future projection, which runs with historical forcing during 1950–2014 and future forcing derived from the SSP5–8.5 scenario during 2015–2050, respectively. We exclude the MPI-ESM1-2-HR because of its poor performance on WNP monthly TC frequency and the failure to capture the intra-year seasonal variation of WNP TC frequency38. ECMWF-IFS-HR is also eliminated due to the lack of Tier 3 (highresSST-future). Therefore, four models are selected for this study: CMCC-CM2-VHR4, CNRM-CM6-1-HR, EC-Earth3P-HR, and HadGEM3-GC31-HM. Two tracking algorithms, TRACK42 and TempestExtremes41 (referred to as TempExt), are applied for TC identification for each HighResMIP model. TRACK tracks features using relative vorticity with criteria for warm-core and lifetime and TempExt is based on tracking sea level pressure (SLP) features with criteria for warm-core and lifetime. TC data are detected from each model in both tracking algorithms58,59, and the multimodel ensemble mean (MME) is calculated separately based on each tracker, referred to as MME_TRACK and MME_TempExt, respectively. Supplementary Table 1 provides specific descriptions of the four models and datasets, among which CMCC-CM2-VHR4 has the highest resolution of 38 km, and HadGEM3-GC31-HM has the lowest resolution of 75 km. The TC genesis definition is the same as in the observation data. If a model has more than one member, the equal-weighted mean of these multimodel ensembles is used. Environmental fields (e.g., horizontal winds, relative humidity, and vertical velocity) interpolated into a 1° × 1° resolution are also multimodel averages. Since the sea surface temperature (SST) is not available in the models, the air surface temperature over the oceans is used as a replacement.

Dynamics Genesis potential index

Vertical wind shear is the magnitude difference between 200 and 850 hPa horizontal winds. In addition, we employ the dynamic genesis potential index (DGPI) developed by Wang and Murakami60 based on dynamic factors and further conduct a relative contribution analysis61,62. The formula is as follows:

$$\begin{array}{ll}{{{DGPI}}}={(2.0+1.0\times {V}_{{\rm {s}}})}^{-1.7}{(5.5-\frac{{\rm{d}}{u}_{500}}{{\rm{d}}y})}^{\!2.3}{(5-20\times {\omega }_{500})}^{3.4}\\{(5.5+\left|{\zeta }_{{\rm {a}}850}\times {10}^{5}\right|)}^{2.4}{{\rm{e}}}^{-11.8}-1.0\end{array}$$
(1)

\({V}_{\rm {{s}}}\) is vertical wind shear (m s−1); \({\zeta }_{{\rm {a}}850}\) is 850-hPa absolute vorticity (10−5 s−1); \({u}_{500}\) is 500-hPa zonal wind (m s−1); and \({\omega }_{500}\) is 500-hPa vertical velocity (Pa s−1).

Metrics and statistical methods

This study compares the muti-year average difference between 1980–2010 and 2020–2050 to examine future changes in TCs and environmental fields over the WNP. The TC peak season is defined from June to October (referred to as JJASO) since the TC frequency over the WNP during these five months accounts for 75% of the annual frequency. Changes in environmental fields (including SST, wind fields, etc.) focus on the TC peak season, and the annual results are similar (not shown). LMI latitude refers to the location where a TC first reaches its maximum intensity over the WNP. The Student’s t-test is used to examine the significance of the changes.